I'd love to see the kind of results you get too (and curious about your comfy workflow too.. Mines are a mess ^^').

Oh, that is a good tip. Makes some really nice results, and seems to also increase the prompt adherence even more. I really like it.

How do you get the new sampler?

@airuru They generally come fixed with most of the UI's commonly used. I and vrgamerz use comfy UI, the samplers are included in that. Other UI's will have some of the same, like the absolute standards like euler ancestral, dpm++ 2m/sde and such, but will and will not have other samplers. For many UI's there also isn't any convinient way to add samplers. Especially when using an online generator like the one on this site. Then you are definitely stuck with what is available.

First thing I would try is adding stuff like "text" and similar to your negative promt. It also helps to check if some of the keywords you used can mean text in an image or have a strong connection to writing or the kind of text poping up in your images and find an alternative way to prompt for it. When I have unexpected results I usually read over my promt and check for that. Oftentimes it is one small word that alone or in context can have a differnt meaning and the model is running with that. Like when I had a whole lot of exaggerated hips on women and didn't know why. In the end it was me using "thick treetrunk" for the backgrund prompt. And changin that solved the problem. Use a thesaurus is a tip I always reccomend. The base is still sdxl. So the danbooru tags while helpful are not the be all end all. You can get pretty neat and interesting results with using less comon words that have similar meanings in your prompts.

@Oybou thank you

I feel the same. Other models I just couldn't get to do certain things I wanted the way this one does. it has by far been my favorite I have used.

Very good analysis. Well done. Need more content like this

@Lja Thank you. I don't comment a lot, but when I do I try to be helpful and informative. There is quite often a glaring lack of information here (on the internet in general really) and if I can I will try to help correcting that and make life easier for other people. Because that is what I would have liked quite often so when my comments help people it makes me happy.

As a further note now that I am already writing, in comfy the er sde sampler (as someone else already mentioned too in the comments) works really well with version 4 (I will be testing with 5 soon a see what changed). It males for very stable results that are usually very close to the prompt. But the downside is that your results will reduce a lot in variation. I is less "creative" but a great choice if one prompts for something specific. For more creative results, euler ancestral is the choice, as always it seems. But dpm++ 3m sde can also give quite interesting results. For a sheduler, I preder sgm uniform or karras. Both make for good results, although I feel uniform does better at hands.

The checkpoint also generally deals well with most lora, but quite a few loras can quickly lead to degradation of finer details. But that is something that happens with most models from my experience.

@Oybou I try to comment regularly but I have not been doing reviews as much or detailed analysis like what you have done. Most people do not comment at all. This is a social norm I hope changes because just merely saying thanks even if it isn't a major contribution is easy to do and is nice for the creators.